High-Dimensional Multi-Source Feature Fusion for Early Default Prediction in Consumer Credit Portfolios

This study develops a multi-source feature-fusion framework that combines transaction histories, mobile-behavior data, credit-bureau information, and merchant-level attributes. The feature space contains over 4,800 engineered variables derived from 3.5 million customer records. A three-stage selection pipeline—correlation filtering, mutual-information ranking, and stability-selection LASSO—reduces dimensionality by 92%. The selected features train a LightGBM model optimized for early-stage (0–30 day) delinquency prediction. The model achieves an ROC-AUC of 0.91 and reduces false-negative early defaults by 37.5% compared with baseline logistic regression. Feature-importance patterns reveal strong interactions between merchant category instability and device-behavior anomalies. The results show the effectiveness of multi-source feature fusion for fine-grained default prediction.